Run computer-use agents inside cross-OS desktop sandboxes with CUA

Launch computer-use agents in Linux, macOS, Windows, or Android sandboxes and drive full desktop tasks through screenshots, mouse, keyboard, and shell controls.

Prerequisites

Python 3.11+ for the SDK or Node.js for cuabot, plus local QEMU or CUA cloud access depending on runtime choice.

Installation

Use the upstream install or setup path that matches your environment:

• pip install cua
• uv tool install -e . && cb image create linux-docker

Requirements and caveats from upstream:

• python

• Requires Python 3.11 or later

• | lumier | Docker-compatible interface for Lume VMs |

Basic usage or getting-started notes:

• result = await sb.shell.run("echo hello")

• cd cua-bench

• Run benchmark with agent

• Source: https://github.com/trycua/cua

• Extracted from upstream docs: https://raw.githubusercontent.com/trycua/cua/HEAD/README.md

Documentation

• https://cua.ai

Source

• Agent Skill Exchange